AI Digest.

Claude Fable 5 Unleashed: Self-Rewriting Agent Loops, 10x Benchmark Gains, and a Flood of New Coding Models

Claude Fable 5 dominated today's AI discourse with demonstrations ranging from self-rewriting agent loops to parametric turbofan CAD, while benchmark results showed a 10x improvement over other frontier models on post-training tasks. Three new coding models launched within hours of each other, and Cognition open-sourced Devin's most popular cloud handoff feature.

Daily Wrap-Up

Today was Fable 5 day. Anthropic's latest model release didn't just make appearances across the feed, it practically took it over. We saw it build a complete Airbus-class turbofan engine in parametric CAD, power interactive Three.js simulations that one user described as having reached a quality threshold where "it is now impossible to explore all possibilities," and drive Firecrawl's new Prometheus agent for automated web data collection. But the most striking moment came from @thoughtfullab's benchmark results shared by @scaling01: Fable 5 spent 17 hours and 25 million tokens training a weaker model to solve a puzzle, achieving 34% pass@1 where every other frontier model averages under 4%. That is not an incremental improvement. It is a qualitative leap that explains why Anthropic appears to have placed restrictions on Fable 5's ability to develop LLMs, as @scaling01 pointedly noted.

The philosophical underpinning tying all this together is what @swyx and Latent Space are calling "Loopcraft": the emerging discipline of stacking self-improving agent loops. The "Salty Lesson" for agents parallels Sutton's Bitter Lesson for models. Don't fix things yourself. Build systems that scale with more agents through better goals and orchestration. This maps directly to @0x_rody's share from Anthropic's event, where a Metaview engineer explained they stopped fixing prompts entirely because "the system reviews its own output and rewrites its own instructions now." Meanwhile the model marketplace continues its relentless compression. NVIDIA's Nemotron 3, MiniMax's M3, Kimi's K2.7-Code, and Qwopus 3.6 all launched or gained attention today, each pushing on inference speed, context length, or cost. The most practical takeaway for developers: stop hand-tuning prompts and start building self-correcting agent loops. The teams seeing outsized results are the ones letting models iterate on their own instructions, not the ones manually crafting perfect system messages.

Quick Hits

  • @Krongggggg highlights a 25-day scaling fundamentals series by @system_monarch, a Principal Engineer at Atlassian, covering load balancing, system design, and architecture patterns from someone who has designed systems handling millions of requests.
  • @BiaNeuroscience pitches Bía Sleep, a neurofeedback wearable built on 40 years of neuroscience that uses real-time brain signals to optimize deep sleep, currently

Sources

B
Bía @BiaNeuroscience ·
Bía Sleep is backed by sleep experts and built on 40 years of neuroscience. By using real-time neurofeedback Bía helps calm a racing mind, optimize deep sleep, and wake you up strong and ready to take on the day. Order Bía Sleep today and receive a free lifetime subscription.
S
Superconductor @Superconductor ·
Launch Claude Code / Codex agents, guide them, and review PRs from your phone.
R
rody @0x_rody ·
At Anthropic's event, Metaview engineer: "We stopped fixing our prompts. The system reviews its own output and rewrites its own instructions now." In 16 minutes, he shows the Claude Code loop running in production on thousands of reviews, not in a demo. Watch the talk, then grab the full loop setup below👇
0 0x_rody @0x_rody

https://t.co/LWQbDMXIdK

D
Daniel @daniel_z1909 ·
🛫 A complete Airbus-class turbofan — fully parametric, animated, built entirely in the browser. Created in confBuild with Claude Fable 5: ⚙️ Real internals — 7-stage compressor, annular combustor, 4 turbine stages 🌀 Two-spool animation: HP & LP shafts at real differential speed 🔥 Flow lines & exhaust react live to the N1 lever 📐 Dimensions on sliders — everything recalculates exactly 📄 Export to STEP, auto-generate technical drawings Every chamfer and bore: calculated, not painted. #CAD #Engineering #Turbofan #Aerospace #confBuild #AI
크롱 @Krongggggg ·
아틀라시안 수석 급이 수백만 요청도 버티는 스케일링 노하우를 무료 연재하는 건 주니어 시니어 불문하고 개이득임. 25일간 스케일링 기초 연재 시작했는데, 1일 차 로드 밸런싱 원리랑 라운드 로빈 알고리즘 부터 정리해 줌. 첫날부터 로드 밸런싱 알고리즘 헛점을 잪어주는데 이런 경험담은 어디가서 들어보기도 힘듬. 일단 팔로우 하트 북마크 하고 본다.
S system_monarch @system_monarch

I've been a backend Engineer for 12+ years. Today, I'm a Principal Engineer at Atlassian. I've designed systems that handle millions of requests. Sat on both sides of system design interviews. Reviewed more architecture docs than I can count. Starting today, I'm breaking down the fundamentals of scaling for the next 25 days. If you're learning system design bookmark this thread, you're going to get a lot of learning from this.

L
Liliana Hotsko @liliana_hotsko ·
How do you give a code LLM knowledge of an entire repository without paying for it at every single query? We introduce Code2LoRA: a hypernetwork that turns a repository into its own LoRA adapter. Repo knowledge baked into weights → zero inference-time token overhead. https://t.co/T3hnxCVkna
G
Gokul R Nair @itIsGokulNair ·
guess what? No more messy pipelines. Fine-tune → quantize → export GGUF. All in one app. https://t.co/iW9LozrvnL #localAI #ondeviceAI #AITools https://t.co/YdnFVWOIM2
F
Firecrawl @firecrawl ·
Introducing Prometheus, an experimental Forward Deployed Agent for web data. Describe the web data you need and it writes Firecrawl code to collect it. Run it yourself or let us host and automatically maintain it as pages change. Try it with Claude Fable 5 for free this week! https://t.co/DF9ZLUNTOX
A
Akshay 🚀 @akshay_pachaar ·
Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. I'll soon publish another article on speculative decoding. Stay tuned!!
A akshay_pachaar @akshay_pachaar

KV Caching in LLMs, Clearly Explained

A
Alvin Sng @alvinsng ·
For those who think their code is secure: it isn't. We've been running automated security scans on every PR and finding countless vulnerabilities every week. Chances are your codebase contains thousands of vulnerabilities, some minor, but some severe enough to cause real damage.
F FactoryAI @FactoryAI

Introducing Automated Security Review in Droid. https://t.co/47ZhOUWdCt

B
Bryce Roberts @bryce ·
1M+ on a product launch tweet. Some of the top builders. Hundreds of leads. Some of the largest builders scheduled. Caleb and team are just getting started.
C calebarclay @calebarclay

We're launching Bridge today 🌉 An AI engine that builds virtual homes. Blueprint in, walkable home out. Every plan, every option, structural changes included. What took 3D artists months now takes days. Homebuilders can finally show buyers every home they sell. https://t.co/QrwlM5FUjP

C
Cognition @cognition ·
/handoff is one of the most powerful features of Devin CLI: close your laptop, your agents keep working from the cloud. Now, we've open sourced it.
I imjaredz @imjaredz

We've open sourced my favorite Devin feature: /handoff Hand off jobs to cloud Devins from your local machine Install it as a plugin in Claude Code or Codex or any other coding agent Close your laptop without pausing your agents 😉 https://t.co/1HMgfzuar6

N
NVIDIA AI @NVIDIAAI ·
Shoutout to Caleb for putting together a great deep dive on Nemotron 3 🙌 Check it out.
C calebfoundry @calebfoundry

Nemotron 3 Full Breakdown With the help of Joey Conway from @NVIDIAAI getting into the specifics around why Nemotron 3 is kind of a big deal Biggest headline with Nemotron is: Hybrid Mamba Transformer, Latent MoE, and MTP Hybrid Mamba Transformer essentially attacks right at the Attention mechanism to make the overhead sub-quadratic, but unlike quantizing KV Cache or swapping out attention head, NVIDIA chose Mamba-2 Latent MoE helps further optimize on sparsity by down projecting the dimensions so you're doing less math and less memory movement between HBM and SRAM, you're saving a ton, and NVIDIA made a conscious choice to add more experts given the surplus Finally, MTP or multi token prediction where the model can see future tokens to be more expressive in training and also option to use for speculative decoding during inference Oh, also the model adopts the new OpenMDW 1.1 License

P
pfung @philfung ·
I'm not the best judge of robotics simulators, but this is easily the most user-friendly one I've used. It actually runs well on a Mac. Feels more like the Unity editor / game engine than other simulators. https://t.co/Y0VD3eWNla" https://t.co/YUixYkeG0W
L luckyrobots @luckyrobots

We are super excited to share with you our initial release of Lucky Engine. We are building a robotics engine from the ground up to be what we wished we could find in a simulator before https://t.co/fR10g5iRXg

L
Lisan al Gaib @scaling01 ·
and this is why Anthropic restricted LLM development for Fable 5
T thoughtfullab @thoughtfullab

Fable 5 is doing something wild on our FrogsGame post-training task. It trains a weaker model to solve the puzzle, peaks at 68%, and produces the only ~10x improvement we see across the benchmark. It spent 17 hours, 25M tokens without human in sight. 34% pass@1, while every other frontier model averages under 4%. We will publish a more detailed analysis soon.

G
Geek Lite @QingQ77 ·
给编码智能体用的 HTML 技能,专门生成简洁的架构图、计划页和视觉文档。 https://t.co/rrYGQlWbYJ effective-html 是一套给智能体用的 HTML 技能,专门做自包含的、好看的 HTML 交付物。三个子技能:html 做通用页面,html-diagram 做全屏架构图和系统图(SVG 优先),html-plan 做计划页。 https://t.co/wBwCd0gJK9
0
0xSero @0xSero ·
Coder bros with 32GB memory
K KyleHessling1 @KyleHessling1

Qwopus 3.6 27b-Coder is now live! Scores a 67% on a full run of SWE bench verified with thinking completely disabled! Q5_K_M This model is lightning fast for dense class! With a natively finetuned MTP head, it achieves 100 tps on a single 5090! The biggest upgrade here, though, is its stability in programming and tool calling within @NousResearch Hermes agent, with thinking off! Wall time is crazy fast this way, which makes Hermes feel "native" and snappy, like they were meant for each other. The freedom of running without thinking at all makes you part of the thinking process, and you never get caught waiting 15 minutes for it to finish a thought string, like with the base models. Thinking on and temp high, .9-1 seems to produce really incredible design and svg results. I reran the Boat survival prompt through a few turns, thinking on, and it seemed to render more fancy models in HTML canvas, but it was much more of a start-a-prompt and wait experience vs the snappy and active iteration with it disabled. It may be worth turning it off and on throughout the build process if you want to get really creative with design. Really looking forward to seeing how this one performs for y'all! Please post comments with your opinions and use cases below! As always with our fine-tunes, mess with the temperature setting, and run them much hotter than the base! Please check out the Boat Survival game I posted yesterday, made in 12 turns using Hermes and this model, with thinking off. Link below! Full swe bench repo-specific breakdown also posted in the comments for those interested! Happy building, everyone! We're looking forward to your thoughts! Quants uploading now! https://t.co/kxJE3C39ZZ

D
Drew Breunig @dbreunig ·
So is everyone going to manage multiple versions of skills? https://t.co/evo9R3OeW3
D
Drew Breunig @dbreunig ·
Btw, if you provide a skill to users and want to ship a Fable version, check out the pattern in gskill:
S ShangyinT @ShangyinT

GEPA for skills is here! Introducing gskill, an automated pipeline to learn agent skills with @gepa_ai. With learned skills, we boost Claude Code’s repository task resolution rate to near-perfect levels, while making it 47% faster. Here's how we did it: https://t.co/VsWMyZncC9

S
swyx @swyx ·
## On Loopcraft One might argue the entire game of the next century is to be able to stack loops as effectively as possible. In the early days of each phase, it will be valuable to know when to go **DOWN** a loop when things go wrong (for reliability)… but it will probably be more valuable to know how to go **UP** a loop as models improve (for leverage). If you don’t figure out how to do this, don’t be salty when you lose to those that do.
L latentspacepod @latentspacepod

[AINews] Loopcraft: The Art of Stacking Loops @RichardSSutton has his “Bitter Lesson” for models. We now have the Salty Lesson for agents: Don’t fix things yourself, as you have done historically. Instead focus on systems that scale with more agents, like goals and orchestration. More in today's op-ed: https://t.co/EyYV4VGRpi

A
Ahmad @TheAhmadOsman ·
Wrote an article on the case against Anthropic as a safety-branded permission regime for Cognition Infrastructure Covers sabotage-as-safety, anti-Opensource rules, Fable, regulatory capture, data asymmetry, Claude Code, and Who Owns Intelligence Must read, bookmark it for later
T TheAhmadOsman @TheAhmadOsman

Anthropic's War on Opensource AI

I
Isra @israfill ·
use minimax M3 for FREE through tokenrouter - 1M context, zero cost 😳 M3 beats GPT-5.5 on coding benchmarks (SWE-Bench Pro: 59% vs 58.6%) what you get for $0: - SWE-Bench Pro 59% (ahead of GPT-5.5, trails opus 4.8 by ~10 pts) - 1M context window actually usable at 500k+ tokens (opus caps at 200k) - 9-15x faster inference at long contexts (sparse attention) - free input AND output literally $0/0 M3's real edge over opus right now: - 1M context for full codebase dumps - speed: fast enough for real-time agent loops - 1/20th the compute cost of opus for equivalent tasks how to set up (3 min): > go to https://t.co/EtW28QHsgI and sign up (email + verify) > open API Key section and create a new key > copy the base URL and the key > paste into any openai-compatible client (hermes, claude code, cursor, https://t.co/5WBKCvaP6h) > base url: https://t.co/fzTiCsLBq2 > model: minimax/M3 important: - no credit card, no phone verify, no trial - one key also routes GPT, claude, gemini, deepseek - rate limits apply on free tier, fine for personal use and agent prototyping - free window ends june 17, 5 days left sota-level coding + 1M context + speed = $0 while everyone else pays $200/mo for claude api bookmark this before the free window closes
I israfill @israfill

use kimi k2.6 for FREE through nvidia's api AND their new desktop app with 300 agents 😳 kimi beats GPT-5 on coding benchmarks what you will get for $0: - kimi k2.6: SWE-Bench Pro 58.6, Multilingual 76.7 - free for 1 year on NVIDIA's API - 4,000+ tool calls per session - 300 parallel agents what the new desktop app does: - runs native agent swarms on mac/windows - builds full PPTX, Word, Excel, PDF files - agents browse the web, click, type, fill forms for you - built-in finance data (yahoo, binance, world bank) - remembers your workflow + turns repeated tasks into skills how to get both for free (5 min): > go to https://t.co/BIIXYD9PJA and sign up (web + desktop) > go to https://t.co/V6J008DsGg, search kimi-k2.6 > generate a free api key > paste into any openai-compatible client (cursor, cline, etc) > base url: https://t.co/snZSjFRN41 important: - nvidia tier has ~40 req/min rate limit - desktop free tier has some agent limits - not for production workloads - phone verify may be needed for nvidia sota coding model + full desktop agent suite = $0 while most people are paying $20/mo for each bookmark this before the free tiers change

E
echo.hive @hive_echo ·
hall of profiles by fable's imaginative interpretation using @threejs pretty creative i say :) download link in comment as well as a link for fable simulation playground https://t.co/Rjs02tcsP2
H hive_echo @hive_echo

experimentally eerie made with fable and @threejs download in comment as well as fable playground link it feels like model capability with fable have hit a point where it is now impossible to explore all possibilities. before the quality just wan't there and frustration was often the end result the realm of possibilities and quality will grow exponentially from here on out is my humble 2 cents

E
echo.hive @hive_echo ·
Play with interactive fable 5 simulations: https://t.co/RDKylk3qwV
N
noname @malikwas1f ·
RT @Kimi_Moonshot: 🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over…